Operational uncertainty in machine-learning based debris block detection in urban waterways

This study investigates the use of machine learning based image classification techniques to detect debris blocking of urban waterways. Using a dataset comprised of 1089 labelled CCTV images of a trash screen located in Cardiff, UK and a comprehensive resampling approach, we investigate not only the ability of selected machine learning algorithms to correctly identify images but also to evaluate the uncertainty of these algorithms conditional on the datasets presented to them. For each candidate model, we considered two datasets: an imbalanced dataset and an undersampled dataset. The results demonstrate that the performance of a simple logistic regression model was broadly comparable to that of more advanced machine learning models such as vision transformers. The best performing models (vision transformers and logistic regression) achieved an accuracy of more than 80% while the NetRes50 model achieved an accuracy in the low 70%. This is an important result that opens the possibility for implementing these techniques as part of an operational real-time flood warning system utilising already existing cameras.